With the increasing amount of scientific publications in digital libraries, it is crucial to capture "deep meta-data" to facilitate more effective search and discovery, like search by topics, research methods, or data sets used in a publication. Such meta-data can also help to better understand and visualize the evolution of research topics or research venues over time. The automatic generation of meaningful deep meta-data from natural-language documents is challenged by the unstructured and often ambiguous nature of publications' content.In this paper, we propose a domain-aware topic modeling technique called Facet Embedding which can generate such deep meta-data in an efficient way. We automatically extract a set of terms according to the key facets relevant to a specific domain (i.e. scientific objective, used data sets, methods, or software, obtained results), relying only on limited manual training. We then cluster and subsume similar facet terms according to their semantic similarity into facet topics. To showcase the effectiveness and performance of our approach, we present the results of a quantitative and qualitative analysis performed on ten different conference series in a Digital Library setting, focusing on the effectiveness for document search, but also for visualizing scientific trends.